Linux Load Average Explained Without Lies

Everyone reads uptime, sees load average: 24.0, and panics about CPU. But load average isn't CPU usage — and misunderstanding it sends people adding cores to a box that's actually starved for disk.

What the three numbers are

load average: 24.74, 18.10, 12.93

Those are the 1-, 5-, and 15-minute averages of the number of processes that are either:

• runnable (using or waiting for CPU), or
• in uninterruptible sleep (state D) — almost always waiting on I/O.

That second category is the part everyone forgets. A process blocked on a slow disk counts toward load even though it's using zero CPU.

The classic trap: high load, low CPU

uptime
# load average: 26.40, 25.11, 22.84

top
# %Cpu(s): 5.0 us, 2.1 sy, 8.6 id, 84.3 wa

CPU is 90% idle, but wa (I/O wait) is 84%. The CPU isn't the bottleneck — the disk is. Confirm it:

iostat -x 1 3
# %util near 100, await in the tens/hundreds of ms = the disk is the wall

ps -eo state,pid,cmd | grep '^D'
# processes stuck in D = blocked on I/O

Adding CPUs here does nothing. The fix is to find the I/O hog (a runaway backup, a verbose log, a failing disk), throttle or reschedule it, and check disk health.

Is a given number "bad"?

Load is relative to how many CPUs you have. A load of 4 on a 4-core box means it's fully utilised; on a 32-core box it's cruising. A rough rule: divide load by core count.

nproc            # how many CPUs

But even "load per core" only tells you about demand, not cause. Always pair it with: is this CPU demand or I/O wait?

A quick triage flow

1. uptime — is load high relative to nproc?
2. top — is it user/sys CPU, or wa (I/O wait)?
3. If CPU: which process? (top, pidstat)
4. If I/O wait: iostat -x, find D-state processes, check the disk.
5. If neither adds up: check memory pressure and swap (free -h, vmstat).

The one-liner to remember

High load + busy CPU = compute problem. High load + idle CPU = I/O problem. Either way, measure before you reach for a fix.

Put it into practice in the puzzle High Load, Low CPU or the Linux Black Box lab.